chore(training): retune stage-4 handling penalties and restart from Stage-3 foundation

Stage 4's upright/forward-motion telemetry plateaued flat across all three
blocked attempts because ground_tilt_penalty (0.003) was too weak to matter
and nothing penalized sideways/reverse motion at all. Raise
ground_tilt_penalty to 0.05 and add a new non_forward_penalty term
(ship_ai_controller.gd) that directly costs non-forward planar velocity near
the floor, independent of the ball. Delete the three blocked attempts'
checkpoints/logs/exports and reset generation5_state.json so the next run
starts fresh from the Stage-3 foundation checkpoint instead of continuing
from the drifted retry2 weights.

Game/scripts/ship_ai_controller.gd

@@ -74,6 +74,11 @@ extends AIController3D
 # pitch/roll during a real aerial. Generation 5 uses this instead of raising
 # the global tilt_penalty back to its pre-flight value.
 @export var ground_tilt_penalty := 0.0
+# Per-tick penalty on the planar-velocity component not pointed along the
+# nose (sideways or reverse), independent of the ball — the mirror image of
+# forward_velocity_to_ball_weight's ball-conditioned bonus. Same
+# GROUND_HANDLING_HEIGHT altitude fade as ground_tilt_penalty.
+@export var non_forward_penalty := 0.0
 # Per-tick bonus for own speed: 0 stationary, full value (+0.24/s) at
 # max_speed. Run07 lesson: after the kickoff flurry both ships parked next to
 # a cornered ball — with every other dense term near zero there, standing
@@ -324,6 +329,20 @@ func _physics_process(delta):
 		var tilt_ground_factor: float = 1.0 - clampf(ship.global_position.y / GROUND_HANDLING_HEIGHT, 0.0, 1.0)
 		reward -= ground_tilt_penalty * (1.0 - ground_uprightness) * 0.5 * tilt_ground_factor
 
+	# Dense penalty: any planar velocity component not pointed along the nose
+	# (sideways or reverse), independent of the ball — the mirror image of
+	# forward_velocity_to_ball_weight's ball-conditioned bonus. Fades out with
+	# altitude via the same GROUND_HANDLING_HEIGHT ramp as ground_tilt_penalty.
+	if non_forward_penalty > 0.0 and ship.global_position.y < GROUND_HANDLING_HEIGHT:
+		var non_forward_planar_velocity := Vector3(ship.linear_velocity.x, 0.0, ship.linear_velocity.z)
+		var non_forward_planar_speed := non_forward_planar_velocity.length()
+		var non_forward_planar_forward := Vector3(-ship.global_transform.basis.z.x, 0.0, -ship.global_transform.basis.z.z)
+		if non_forward_planar_speed > 0.0001 and non_forward_planar_forward.length_squared() > 0.0001:
+			var forward_component: float = maxf(non_forward_planar_velocity.dot(non_forward_planar_forward.normalized()), 0.0)
+			var non_forward_speed: float = non_forward_planar_speed - forward_component
+			var non_forward_ground_factor: float = 1.0 - clampf(ship.global_position.y / GROUND_HANDLING_HEIGHT, 0.0, 1.0)
+			reward -= non_forward_penalty * (non_forward_speed / ship.max_speed) * non_forward_ground_factor
+
 	# Dense penalty: height above the floor (see airborne_penalty). The
 	# floor sits at world y = 0 (see training_mode.gd's FIELD_MIN_Y/
 	# _escaped bounds); normalized so the worst case is pinned at the

Game/scripts/training_mode.gd

@@ -251,7 +251,7 @@ const SHIP_AI_OVERRIDES := [
 	"ball_touch_reward", "ball_touch_cooldown_ticks", "ball_touch_direction_floor",
 	"velocity_to_ball_weight", "ball_velocity_to_goal_weight", "ball_distance_penalty",
 	"forward_velocity_to_ball_weight", "wall_contact_penalty", "tilt_penalty",
-	"ground_tilt_penalty", "speed_reward_weight", "time_penalty",
+	"ground_tilt_penalty", "non_forward_penalty", "speed_reward_weight", "time_penalty",
 	"airborne_penalty",
 ]
 
@@ -312,6 +312,7 @@ func _ai_default(name: String) -> Variant:
 		"wall_contact_penalty": return 0.0025
 		"tilt_penalty": return 0.0005
 		"ground_tilt_penalty": return 0.0
+		"non_forward_penalty": return 0.0
 		"speed_reward_weight": return 0.004
 		"time_penalty": return 0.001
 		"airborne_penalty": return 0.0